Limitations of gravity models in predicting fine-scale spatial-temporal urban mobility networks

Abstract

Abstract This study identifies the limitations and underlying characteristics of urban mobility networks that influence the performance of the gravity model. The gravity model is a widely-used approach for estimating and predicting population flows in urban mobility networks. Prior studies have reported very good performance results for the gravity model as they tested it using origin-destination (O-D) data at certain levels of aggregation. Also, the main premise of the gravity model in urban networks is the existence of the scale-free property. The characteristics of urban mobility networks, such as scale-free properties, network size, the existence of hubs and giant components, however, might vary depending on the spatial and temporal resolutions of data based on which these networks are constructed. Hence, the sensitivity of gravity model performance to variation in the level of aggregation of data and the temporal and spatial scale of urban mobility networks needs to be examined. To address this gap, this study examined the basic gravity model, which captures the flow magnitude between O-D pairs based on three variables (population of the origin, population of the destination, and the distance between them). Accordingly, we constructed the urban mobility networks using fine-grained location-based human mobility data for multiple US metropolitan counties. The constructed urban mobility networks have finer resolution as they capture population flow among census tracts on an hourly and daily scale (as opposed to previous studies which used larger spatial blocks). The results show that the scale-free property does not always exist when urban mobility networks are constructed from data with finer spatial and temporal resolution. By examining the association between macroscopic network characteristics, such as the number of nodes and links, average degree, average clustering coefficient, assortativity coefficient, and predictive performance, we found weak association between performance and certain network structures. The findings suggest that: (1) finer-scale urban mobility networks do not demonstrate a scale-free property; (2) the performance of the basic gravity model decays for predicting population flow in the finer-scale urban mobility networks; (3) the variations in population density distribution and mobility network structure and properties across counties do not significantly influence the performance of gravity models. Hence, gravity models may not be suitable for modeling urban mobility networks with daily or hourly aggregation of census tract to census tract movements. The findings highlight the need for new-generation urban mobility network models or machine learning approaches to better predict fine-scale and high temporal-resolution urban mobility networks.